Material proportion guided hyperspectral image super-resolution via unmixing

Existing hyperspectral image super-resolution methods tend to increase spatial resolution subject to human visual preferences, ignoring the accurate recovery of material proportion information. To address this deficiency, we develop a novel hyperspectral image super-resolution framework guided by material proportion priors. This framework is the first to incorporate material proportion information into hyperspectral single-image super-resolution. This framework comprises two models, i.e., a material proportion guidance model and a super-resolution model. The material proportion guidance model takes a low-resolution hyperspectral image as input and produces a feature representation that incorporates material proportion information, which enables the guidance of accurate super-resolution. Specifically, this model consists of three networks, i.e., an unmixing network for material proportion estimation, a rectification network for error correction, and a reconstruction network for material proportion guidance feature generation. In parallel, the super-resolution model is primarily implemented via a super-resolution network that enhances spatial details by super-resolving the same low-resolution hyperspectral image into a refined spatial structural representation. Lastly, the outputs of the two models are fused to generate a high-resolution image containing accurate material proportion information. Comprehensive evaluations conducted on the Houston and QUST-1 Qingdao datasets confirm that our framework achieves superior performance compared with existing state-of-the-art techniques, both in quantitative metrics and in reconstructed visual fidelity under diverse settings. The source code has been made publicly accessible at https://gitee.com/xu_yinghao/MaterialSR.